|
| 1 | +// Copyright (c) Microsoft Corporation. All rights reserved. |
| 2 | +// Licensed under the MIT License. |
| 3 | + |
| 4 | +#include "precomp.h" |
| 5 | + |
| 6 | +namespace Dml |
| 7 | +{ |
| 8 | + |
| 9 | +class DmlOperatorSkipLayerNormalization : public DmlOperator |
| 10 | +{ |
| 11 | +public: |
| 12 | + DmlOperatorSkipLayerNormalization(const MLOperatorKernelCreationContext& kernelCreationContext) |
| 13 | + : DmlOperator(kernelCreationContext) |
| 14 | + { |
| 15 | + std::vector<std::optional<uint32_t>> kernelInputIndices = {0, 1, 2, 3, 4}; |
| 16 | + |
| 17 | + DmlOperator::Initialize( |
| 18 | + kernelCreationContext, |
| 19 | + kernelInputIndices, |
| 20 | + std::nullopt, |
| 21 | + kernelCreationContext.GetTensorShapeDescription().GetInputTensorShape(0), |
| 22 | + std::nullopt, |
| 23 | + kernelCreationContext.GetTensorShapeDescription().GetInputTensorDimensionCount(0)); |
| 24 | + |
| 25 | + const float epsilon = kernelCreationContext.GetOptionalAttribute<float>(AttrName::Epsilon, DefaultEpsilon); |
| 26 | + |
| 27 | + int32_t onnxAxis = kernelCreationContext.GetOptionalAttribute<int32_t>(AttrName::Axis, -1); |
| 28 | + uint32_t inputDimCount = kernelCreationContext.GetTensorShapeDescription().GetInputTensorDimensionCount(0); |
| 29 | + onnxAxis = OperatorHelper::HandleNegativeAxis(onnxAxis, inputDimCount); |
| 30 | + std::vector<uint32_t> onnxAxes(inputDimCount - onnxAxis); |
| 31 | + std::iota(onnxAxes.begin(), onnxAxes.end(), onnxAxis); |
| 32 | + |
| 33 | + assert(m_inputTensorDescs.size() == 5); |
| 34 | + assert(m_outputTensorDescs.size() == 1); |
| 35 | + |
| 36 | + auto inputDesc = m_inputTensorDescs[0].GetDmlDesc(); |
| 37 | + auto skipDesc = m_inputTensorDescs[1].GetDmlDesc(); |
| 38 | + auto gammaDesc = m_inputTensorDescs[2].GetDmlDesc(); |
| 39 | + auto betaDesc = m_inputTensorDescs[3].GetDmlDesc(); |
| 40 | + auto biasDesc = m_inputTensorDescs[4].GetDmlDesc(); |
| 41 | + auto outputDesc = m_outputTensorDescs[0].GetDmlDesc(); |
| 42 | + |
| 43 | + TensorDesc inputSkipBiasTensorDesc(m_inputTensorDescs[0].GetDmlDataType(), m_inputTensorDescs[0].GetSizes()); |
| 44 | + DML_TENSOR_DESC inputSkipBiasDmlTensorDesc = inputSkipBiasTensorDesc.GetDmlDesc(); |
| 45 | + |
| 46 | + DML_ELEMENT_WISE_ADD_OPERATOR_DESC inputSkipAddDesc = {}; |
| 47 | + inputSkipAddDesc.ATensor = &inputDesc; |
| 48 | + inputSkipAddDesc.BTensor = &skipDesc; |
| 49 | + inputSkipAddDesc.OutputTensor = &inputSkipBiasDmlTensorDesc; |
| 50 | + DML_OPERATOR_DESC inputSkipAddOpDesc = { DML_OPERATOR_ELEMENT_WISE_ADD, &inputSkipAddDesc }; |
| 51 | + |
| 52 | + DML_ELEMENT_WISE_ADD_OPERATOR_DESC inputSkipBiasAddDesc = {}; |
| 53 | + inputSkipBiasAddDesc.ATensor = &inputSkipBiasDmlTensorDesc; |
| 54 | + inputSkipBiasAddDesc.BTensor = &biasDesc; |
| 55 | + inputSkipBiasAddDesc.OutputTensor = &inputSkipBiasDmlTensorDesc; |
| 56 | + DML_OPERATOR_DESC inputSkipBiasAddOpDesc = { DML_OPERATOR_ELEMENT_WISE_ADD, &inputSkipBiasAddDesc }; |
| 57 | + |
| 58 | + DML_MEAN_VARIANCE_NORMALIZATION1_OPERATOR_DESC mvnDesc = {}; |
| 59 | + mvnDesc.InputTensor = &inputSkipBiasDmlTensorDesc; |
| 60 | + mvnDesc.ScaleTensor = &gammaDesc; |
| 61 | + mvnDesc.BiasTensor = betaDesc.Desc ? &betaDesc : nullptr; |
| 62 | + mvnDesc.OutputTensor = &outputDesc; |
| 63 | + mvnDesc.Axes = onnxAxes.data(); |
| 64 | + mvnDesc.AxisCount = gsl::narrow_cast<uint32_t>(onnxAxes.size()); |
| 65 | + mvnDesc.NormalizeVariance = true; |
| 66 | + mvnDesc.Epsilon = epsilon; |
| 67 | + mvnDesc.FusedActivation = nullptr; |
| 68 | + DML_OPERATOR_DESC mvnOpDesc = { DML_OPERATOR_MEAN_VARIANCE_NORMALIZATION1, &mvnDesc }; |
| 69 | + |
| 70 | + // Construct the graph |
| 71 | + std::vector<const DML_OPERATOR_DESC*> opDescs; |
| 72 | + opDescs.reserve(3); |
| 73 | + |
| 74 | + std::vector<DML_INPUT_GRAPH_EDGE_DESC> inputEdges; |
| 75 | + inputEdges.reserve(5); |
| 76 | + |
| 77 | + std::vector<DML_INTERMEDIATE_GRAPH_EDGE_DESC> intermediateEdges; |
| 78 | + intermediateEdges.reserve(2); |
| 79 | + |
| 80 | + std::vector<DML_OUTPUT_GRAPH_EDGE_DESC> outputEdges; |
| 81 | + outputEdges.reserve(1); |
| 82 | + |
| 83 | + // Insert the Input + Skip operation into the graph |
| 84 | + opDescs.push_back(&inputSkipAddOpDesc); |
| 85 | + |
| 86 | + DML_INPUT_GRAPH_EDGE_DESC dataInputEdge = {}; |
| 87 | + dataInputEdge.GraphInputIndex = 0; |
| 88 | + dataInputEdge.ToNodeIndex = 0; |
| 89 | + dataInputEdge.ToNodeInputIndex = 0; |
| 90 | + inputEdges.push_back(std::move(dataInputEdge)); |
| 91 | + |
| 92 | + DML_INPUT_GRAPH_EDGE_DESC skipInputEdge = {}; |
| 93 | + skipInputEdge.GraphInputIndex = 1; |
| 94 | + skipInputEdge.ToNodeIndex = 0; |
| 95 | + skipInputEdge.ToNodeInputIndex = 1; |
| 96 | + inputEdges.push_back(std::move(skipInputEdge)); |
| 97 | + |
| 98 | + // Insert the InputSkip + Bias operation into the graph |
| 99 | + if (biasDesc.Desc) |
| 100 | + { |
| 101 | + opDescs.push_back(&inputSkipBiasAddOpDesc); |
| 102 | + |
| 103 | + DML_INTERMEDIATE_GRAPH_EDGE_DESC intermediateEdge = {}; |
| 104 | + intermediateEdge.FromNodeIndex = 0; |
| 105 | + intermediateEdge.FromNodeOutputIndex = 0; |
| 106 | + intermediateEdge.ToNodeIndex = 1; |
| 107 | + intermediateEdge.ToNodeInputIndex = 0; |
| 108 | + intermediateEdges.push_back(std::move(intermediateEdge)); |
| 109 | + |
| 110 | + DML_INPUT_GRAPH_EDGE_DESC biasInputEdge = {}; |
| 111 | + biasInputEdge.GraphInputIndex = 4; |
| 112 | + biasInputEdge.ToNodeIndex = 1; |
| 113 | + biasInputEdge.ToNodeInputIndex = 1; |
| 114 | + inputEdges.push_back(std::move(biasInputEdge)); |
| 115 | + } |
| 116 | + |
| 117 | + // Insert the MVN operation into the graph |
| 118 | + opDescs.push_back(&mvnOpDesc); |
| 119 | + |
| 120 | + DML_INTERMEDIATE_GRAPH_EDGE_DESC intermediateEdge = {}; |
| 121 | + intermediateEdge.FromNodeIndex = biasDesc.Desc ? 1 : 0; |
| 122 | + intermediateEdge.FromNodeOutputIndex = 0; |
| 123 | + intermediateEdge.ToNodeIndex = biasDesc.Desc ? 2 : 1; |
| 124 | + intermediateEdge.ToNodeInputIndex = 0; |
| 125 | + intermediateEdges.push_back(std::move(intermediateEdge)); |
| 126 | + |
| 127 | + DML_INPUT_GRAPH_EDGE_DESC gammaInputEdge = {}; |
| 128 | + gammaInputEdge.GraphInputIndex = 2; |
| 129 | + gammaInputEdge.ToNodeIndex = biasDesc.Desc ? 2 : 1; |
| 130 | + gammaInputEdge.ToNodeInputIndex = 1; |
| 131 | + inputEdges.push_back(std::move(gammaInputEdge)); |
| 132 | + |
| 133 | + if (betaDesc.Desc) |
| 134 | + { |
| 135 | + DML_INPUT_GRAPH_EDGE_DESC betaInputEdge = {}; |
| 136 | + betaInputEdge.GraphInputIndex = 3; |
| 137 | + betaInputEdge.ToNodeIndex = biasDesc.Desc ? 2 : 1; |
| 138 | + betaInputEdge.ToNodeInputIndex = 2; |
| 139 | + inputEdges.push_back(std::move(betaInputEdge)); |
| 140 | + } |
| 141 | + |
| 142 | + DML_OUTPUT_GRAPH_EDGE_DESC outputEdge = {}; |
| 143 | + outputEdge.GraphOutputIndex = 0; |
| 144 | + outputEdge.FromNodeIndex = biasDesc.Desc ? 2 : 1; |
| 145 | + outputEdge.FromNodeOutputIndex = 0; |
| 146 | + outputEdges.push_back(std::move(outputEdge)); |
| 147 | + |
| 148 | + MLOperatorGraphDesc operatorGraphDesc = {}; |
| 149 | + operatorGraphDesc.inputEdgeCount = gsl::narrow_cast<uint32_t>(inputEdges.size()); |
| 150 | + operatorGraphDesc.inputEdges = inputEdges.data(); |
| 151 | + operatorGraphDesc.intermediateEdgeCount = gsl::narrow_cast<uint32_t>(intermediateEdges.size()); |
| 152 | + operatorGraphDesc.intermediateEdges = intermediateEdges.data(); |
| 153 | + operatorGraphDesc.outputEdgeCount = gsl::narrow_cast<uint32_t>(outputEdges.size()); |
| 154 | + operatorGraphDesc.outputEdges = outputEdges.data(); |
| 155 | + operatorGraphDesc.nodeCount = gsl::narrow_cast<uint32_t>(opDescs.size()); |
| 156 | + operatorGraphDesc.nodesAsOpDesc = opDescs.data(); |
| 157 | + |
| 158 | + SetDmlOperatorGraphDesc(std::move(operatorGraphDesc), kernelCreationContext); |
| 159 | + } |
| 160 | +}; |
| 161 | + |
| 162 | +DML_OP_DEFINE_CREATION_FUNCTION(SkipLayerNormalization, DmlOperatorSkipLayerNormalization); |
| 163 | + |
| 164 | +} // namespace Dml |
0 commit comments